基于Metropolis准则的BP模型改进及在太湖Chl-a预测中的应用

水体富营养化及藻华暴发已成为湖泊治理中的主要问题,利用历史监测数据,采用BP神经网络对水体中叶绿素a(Chl-a)浓度进行预测,已成为藻华预警的主要手段.但该方法存在迭代速度慢、易陷入局部极值等局限性,导致产生拟合结果不优或预测误差较大等问题.利用Metropolis接受准则的全局寻优能力,将其与BP神经网络相结合构建...

IMPROVEMENT OF BP MODEL BASED ON METROPOLIS CRITERION AND ITS APPLICATION IN CHLOROPHYLL-A PREDICTION FOR LAKE TAIHU

Eutrophication and algal blooms have become the main problems in lake management. Predicting chlorophyll-a concentration in water based on historical monitoring data using BP neural network is one of the main means for algal bloom early warning. However, traditional BP method has some limitations, such as low iteration speed and being easy to fall into local extremum, leading to a poor fitting result and larger prediction error. In this paper, a new model(MBP) was developed based on BP neural network by combining with the global optimization capability of Metropolis acceptance criterion, and then applied to predict the monthly average chlorophyll-a concentration of Lake Taihu.Resultsshowed that, comparing with traditional BP neural network, the improved MBP model had a relatively faster coverage velocity at the initial iteration stage, and showed lower fitting error and higher accuracy. Average prediction error of the MBP model was significantly lower than that of the traditional BP neural network. Additionally, the MBP model had a stronger robustness and stability for different data noise and smaller number of samples. This model further expanded the application of traditional BP neural network in predicting concentration of chlorophyll-a and provided a new idea for establishing an early warning system of algal bloom.